Why is there such a gap between AI hype and AI results?
There's a gap because the people setting expectations aren't the people doing the work. Leaders forecast transformation; leasing and asset teams see marginal change. McKinsey's The State of AI in 2024 found 65% of organizations regularly use generative AI, yet most report no material bottom-line impact yet. The distance between "we use it" and "it changed our numbers" is where firms stall.
A recent issue of the Legal Tech Trends newsletter (issue #47) led with the phrase "striking divergence in AI expectations." It's a useful label. Legal and CRE are different businesses, but they share a shape: a document-heavy, judgment-heavy operating firm where a few partners are optimistic, a wider bench is skeptical, and the tooling in between rarely earns trust.
A quick definition. By expectations gap we mean the distance between what leadership predicts AI will do and what the operating team actually experiences week to week. It is not measured in demos. It is measured in whether the rent roll is more accurate, whether broker updates stop getting lost, and whether anyone trusts the output enough to send it.
What does the "divergence in expectations" actually look like?
It looks like two people describing the same tool and disagreeing about whether it works. That's not rhetoric — it's what the survey data shows when you split respondents by role and by outcome.
- The measurable version. McKinsey's State of AI in 2024 reports that 65% of organizations regularly use gen AI — roughly double the prior year — while the share reporting meaningful enterprise-wide EBIT impact from it stays in the single digits for most functions. High usage, thin attributed results: that is the divergence, quantified.
- The role version. In practice the optimist is usually a principal who saw a good demo. The skeptic is the leasing coordinator who got a lease abstraction back with the wrong renewal date and now double-checks everything by hand. Both are right about their own experience.
- The legal parallel. The Legal Tech Trends issue #47 framing — "striking divergence in AI expectations" — describes the same split inside law firms: leadership enthusiasm running ahead of front-line trust.
The lesson for CRE: a firm-wide "AI is working" or "AI is useless" verdict is almost always an average of two very different jobs.
Should CRE operators copy legal's "20% time for AI" idea?
Partly. Structured experimentation time is a real, defensible practice — the Legal Tech Trends issue #47 highlighted firms carving out dedicated hours for AI experimentation. Copy the habit of protected time. Do not copy the assumption that free exploration alone closes the gap. Unstructured play widens the trust problem when outputs touch real deals.
If you run something like it, put a fence around it:
- Pick one artifact, not a category. Not "AI for leasing." Pick the leasing tracker, the broker update, or the lease abstraction — one named workflow with a clear owner.
- Define what "right" looks like before you start. A dated renewal, a correct RSF (rentable square footage), an accurate NNN (triple-net) expense figure. If you can't check it, you can't trust it.
- Keep a human on the send. Experimentation drafts; a person approves anything that leaves the building.
What should a CRE operator do differently?
Adopt one governance rule before you adopt any tool: the software proposes, a human commits. Every drafted action — a broker follow-up, a rent-roll update, a lease abstraction — should be staged for review and land on an audit trail, never fired automatically. Vantrow calls this "propose, never commit," but the rule stands on its own regardless of vendor.
Here's why it closes the gap rather than papering over it:
- It converts the skeptic. The leasing coordinator who got burned by a wrong date will trust a system that shows its work and waits for approval — because the cost of a mistake drops to "catch it in review."
- It gives the optimist something real to point to. Instead of a demo, the principal gets a logged trail of drafts approved, edited, or rejected. That's evidence, not enthusiasm.
- It makes the first workflow winnable. Start where the "right answer" is checkable and the blast radius is small. A lease-expiration tracker beats "an AI assistant for the whole firm" every time.
Governed AI — a defined term worth keeping — means the system stages actions inside limits you set, and a human approves before anything commits. That is different from an autonomous agent that acts on its own and asks forgiveness later.
How do you pick the first workflow to ship?
Pick the one where a wrong answer is cheap to catch and expensive to miss. Lease expirations, rent-roll updates, and broker follow-ups all qualify: the output is checkable, the owner is obvious, and a human can approve the send. Avoid open-ended "assistant" projects — they have no checkable output and no clear owner, so trust never forms.
A simple test: can you write down, in one sentence, what "correct" looks like for this task? If yes, ship it. If no, you've picked a category, not a workflow.
FAQ
FAQ
Common questions from operators closing the AI expectations gap.
Q: Should CRE operators run "20% time" for AI experimentation? Yes, but with a fence around it. Protected time to experiment is worth copying from legal firms. Free-for-all exploration is not. Point the time at one named workflow with a checkable output — a lease abstraction, a rent-roll update — and keep a human approving anything that leaves the building.
Q: What does "propose, never commit" mean? It's a governance rule: software drafts an action and stages it for review, but a human approves before anything commits, and every step lands on an audit trail. It lowers the cost of a mistake to "catch it in review," which is what turns a skeptical leasing team into users.
Q: How do you pick the first AI workflow to ship? Pick the one where "correct" is easy to define and a wrong answer is cheap to catch. Lease expirations, rent rolls, and broker follow-ups qualify. If you can't write down what a right answer looks like in one sentence, you've chosen a category, not a workflow.
Q: Why do so many CRE AI pilots stall? Because usage and results are not the same thing. McKinsey's State of AI in 2024 found 65% of organizations use gen AI while few attribute meaningful profit impact to it. Pilots that chase broad "assistant" goals produce no checkable output, so front-line trust never forms and the tool gets abandoned.
Q: What's the difference between governed AI and an autonomous agent? Governed AI stages actions inside limits you set and waits for human approval. An autonomous agent acts on its own and reports afterward. For a document- and judgment-heavy firm handling rent rolls and leases, governed keeps the human on the send — which is exactly where trust and liability live.